A cross-linguistic quantitative analysis of dependence between phonological parameteres
Lorna Rozelle
University of Washington
Interesting quantitative linguistic statements often appear in published papers and are used to
support theoretical arguments. However, these statements are sometimes not verified
quantitatively, let alone statistically, and verifying them can be difficult. This paper presents part
of a quantitative cross-linguistic study of phonological properties of the sign language lexicon.
The ultimate goal of this research is to describe the structure of the lexicons of natural sign
languages; the more immediate goal is to describe systemic properties and patterns in the
lexicons of four languages. The independence of the phonological parameters of handshape,
location, and the number of active hands used a sign’s articulation is investigated here. Do the
values of these parameters occur independently of each other, or does the occurrence of a certain
value for a parameter give information about the value of another parameter? In particular, the
following three questions are addressed:
1.
2.
3.
Are location and number of active hands independent variables, or are certain locations
more likely to host one-handed signs versus two-handed signs?
Are handshape and the number of active hands independent variables, or are certain
handshapes more likely to be used in one-handed signs versus two-handed signs?
Are handshape and location independent variables, or are certain handshapes more likely to
co-occur with certain locations?
Four unrelated and geographically diverse sign languages were investigated: American Sign
Language (ASL), Korean Sign Language (KSL), New Zealand Sign Language (NZSL), and
Finnish Sign Language (Suomalaisen Viittomakieli: SVK). For each language a database was
created that contained approximately 600 signs randomly chosen from published dictionaries.
For each sign, phonological parameters were phonetically transcribed.
To evaluate dependence between variables, two different methods were used. The chi-square test
for correlation is a standard statistical test for determining with what significance it can be
asserted that the observed frequency of occurrence differs from the expected. This test cannot be
used when cell counts are below five. For such cases an information theory test was developed.
The mutual information (MI) between two variables measures how much information is known
about one variable when the value of another variable is known. A nonzero MI can be spuriously
generated by large, sparse tables, such as the handshape x location tables. To determine whether
such a nonzero MI is significant, thus indicating dependence, a computer program, the MI
Significance Program (MISP), was written to compute the MI of 1,000 tables that were created
by randomly scrambling the pairings of the original data. Any dependence that might have
existed in the original data is thus destroyed, and the original MI is compared to the MI of the
scrambled tables to determine its significance. (For data that can be tested either way, the two
methods yield extremely close p values.)
When two variables were found to be dependent by the chi-square test or the MISP, the
following method was used to try to determine the source of the dependence. Handshapes and
locations were grouped into naturally motivated sets, and these sets were tested to see if they
were relevant to the dependency. Handshapes were classified as marked or unmarked. Two
different markedness criteria were used, one language-specific (the handshapes allowed on the
nondominant hand in Type 3 signs) and one language-universal (the six most common
handshapes, accounting for 50% of the data pooled from all four languages). Locations were
grouped as contact or neutral space, and within the contact group as neck/face or torso.
The independence testing results are show in (1). For two variables to be considered dependent,
it must be that p<0.050.
(1) Dependence between the variables of location, handshape and number of active hands
loc x {1 or 2}
hs x {1 or 2}
hs x loc
ASL
p=0.000
p=0.075
p=0.054
KSL
p=0.000
p=0.000
p=0.066
NZSL
p=0.000
p=0.008
p=0.013
SVK
p=0.000
p=0.082
p=0.001
For example, all four languages show dependence between location and the number of active
hands. When locations are classified as contact or neutral space, all four languages again show
dependence, with p=0.000. In all four languages, more one-handed signs contact a location than
expected, while more two-handed signs are articulated in neutral space than expected. When
contact locations are classified as face/neck or torso, ASL, NZSL, and SVK are found to have a
dependency between these variables, with more one-handed signs appearing on the face
(p<0.050). This result is in accord with observation that historically in ASL two-handed signs on
the face have become one-handed (Frishberg 1975). In contrast, these two variables are
independent in KSL, with p=0.489; thus, knowing that a KSL sign is one-handed conveys no
information about whether it is articulated on the face and neck.
Handshape and location are dependent variables only in NZSL and SVK. When locations are
classified as neutral space or contact, the MISP shows that handshape use in NZSL depends on
whether the location is contacted or is in neutral space (p=0.015), while in SVK handshapes are
distributed without regard for this distinction (p=0.457). Then what is the source of the
dependence between handshape and location in SVK? It has been claimed that ASL signs on the
face and neck have more marked handshapes than signs on the torso, because sign perceivers
focus on the signer’s face, and visual acuity is best in the center of focus (Siple 1973; Battison
1995). This claim was tested using both the universal and the language-specific markedness
criteria. The results are shown in (2). Note that the only dependence occurs in SVK, and only
with the universal markedness criterion. There was no dependence in ASL and KSL, which is
unsurprising given the result in (1).
(2) Handshape x Location: marked or unmarked x face/neck or torso
universal
language-specific
ASL
p=0.128
p=0.466
KSL
p=0.138
p=0.946
NZSL
p=0.913
p=0.808
SVK
p=0.018
p=0.115
There are two notable observations. First, patterns of dependence vary greatly crosslinguistically. A natural grouping, such as face/neck versus torso, which strongly influences
whether a sign will be one-handed or two-handed in one language, as in ASL (p=0.001), has no
effect in another language, as in KSL (p=0.498). Second, the concept of markedness, whether it
is defined by a universal, frequency-based criterion or by a language-specific, phonology-based
criterion, is not very useful in accounting for dependence relations between phonological
parameters.